"The challenge is in how you prepare the data and train your AI assistant"
Tine Koloini, CEO of Play123, on building PlayBuddy, a personalised game discovery assistant.
Hello and welcome to AI Gamechangers. Each edition features a candid conversation with someone in the games industry putting AI to work in creative or practical ways.
This week, we're exploring how AI can transform the way players discover casual games. We check in with Slovenia-based Tine Koloini, CEO of Play123, a platform reaching over a million players a month. He tells us how his team deployed a charming AI assistant that’s changing how players engage with browser games, and what the process revealed about prompt design, training data, and player behaviour.
Scroll to the end for your usual dose of curious and interesting AI-related news from around the games industry in the last few days.
Tine Koloini, Play123
Meet Tine Koloini, CEO of Play123, a multilingual browser games platform with a library of 9,000 titles. Drawing on decades of digital marketing and entrepreneurial experience, Tine recently led the addition of PlayBuddy, a ChatGPT-powered assistant that acts more like a knowledgeable friend than a typical support bot.
In this Q&A, he shares what goes into building an AI assistant that adds value, the surprising ways players interact with it, and why prompt design is now core IP.
Top takeaways from this conversation:
PlayBuddy helps users discover and jump into games by interpreting natural-language requests and player moods. User conversations capture insights that regular analytics don’t – from frustration points to unmet content demand.
The biggest challenge in building AI assistants isn't the basic tech: it's data preparation and prompt engineering. Getting the balance between personality and accuracy requires extensive testing with small, incremental changes.
Users engage with AI assistants in unexpected ways, from requesting direct game links to seeking casual conversation after long days, showing the potential for AI as both utility and companionship.
The future points towards hyper-personalised gaming experiences where AI adapts in real-time based on individual patterns.
AI Gamechangers: Let’s start with your origin story. What’s your background, and how long have you been working in games?
Tine Koloini: When I was a kid, my country was part of the socialist bloc, so access to Western technology was limited. My father was a university professor who managed to smuggle home a Commodore 64 following a visit to colleagues in Germany. It came with a small programming guidebook, and so I began experimenting with coding. I was 10 years old at the time. Ever since, I've always been surrounded by computers and games.
“Integrating AI into websites is more than just a new feature. It’s a new way of thinking about how platforms interact with their user base. Gaming has always been about creating memorable player experiences, and combining technology with empathy delivers that”
Tine Koloini
Today, my passion is building products that people genuinely love using. The more they use them, the more fun it becomes for me and my team – it’s what drives us to improve every day. AI fits perfectly into this story. We’re a small team, but with AI, we can do much more, exploring new ideas and trying out different approaches.
Please tell us a little bit about your business. What defines Play123?
Play123 is an online gaming platform that offers a wide selection of free, browser-based games across various categories like puzzles, match-3, and classic casual titles. It’s designed for accessibility and a seamless gameplay experience by allowing users to play instantly on any device without downloads.
Our extensive library is constantly updated with new titles, and we are dedicated to delivering an engaging and social gaming experience. Today, the Play123 network unites 25 domains in as many languages and attracts over a million users each month.
What exactly is PlayBuddy, and how does it work for the typical user on your games platform?
PlayBuddy is an AI-driven assistant that helps users find games based on what they like. Users turn to it when they're trying to find a game but can't remember the name. With PB, they can quickly describe the game they’re looking for, like “game where the vet touches animals and the animal laughs,” and PB will figure out which game they mean.
Because PB has such a friendly personality, users keep coming back for more. Often, they ask PB for suggestions on what to play next. Players describe the mood they are in, and PlayBuddy responds with personalised recommendations, highlighting game features that match their interests.
We designed a conversational system that is efficient and empathetic. PB is like a kind, smart friend who knows all 9,000 games on the platform and is well-informed about gaming in general. PB is always ready to help, no matter your mood. The experience is completely different from standard website support bots that just repeat FAQ answers. PlayBuddy’s focus is on delivering personalised answers in real time, based on the player’s preferences and past gaming history.
PlayBuddy is built using ChatGPT’s API via the Retoba service, correct? Please talk us through how you chose that model and how you worked with them to build it.
We currently use ChatGPT's API as the backend because it consistently provides the most reliable and accurate responses. We are open to exploring other LLMs, and we will switch if we discover a better alternative. On top of that, we use Retoba’s prompt engineering to interact with the API and connect the user and the LLM.
“Traditional analytics don’t reveal users’ intimate thoughts and preferences, but AI interactions do. What content do users want that we currently lack? Which games are frustrating them? These are all questions that AI interactions can answer”
Tine Koloini
I think it is a common misconception that building an AI assistant is primarily a technological challenge. It isn’t. Many companies are providing plug-and-play chatbots these days, so anyone can get started with a few clicks.
The challenge isn’t in the tech, but the data. It matters how you prepare it and how you train your assistant. Most important of all, getting the prompt right is what makes an AI assistant.
What’s your approach to training and fine-tuning? How do you train it with your own internal data?
The first challenge for any company working with AI is how to prepare the data effectively. Which data is relevant? How should we select it, format it, and standardise it? By structuring the data that our internal systems share with the LLM, we optimise the interaction between the two. We now have an automated workflow that feeds the LLM with new game data each week, reflecting the daily changes to our catalogue as we publish new content.
“We believe prompts represent the intellectual property of the future and that prompt engineering will be a highly sought-after profession. Creating an effective prompt comes with many challenges”
Tine Koloini
The hardest part was making the AI assistant act like a real buddy. Sometimes it was overly stubborn; other times, too relaxed.
You’ve mentioned that prompt design is one of the hardest parts of building PlayBuddy. Can you walk us through how you approached it and what the biggest challenges were?
We believe prompts represent the intellectual property of the future and that prompt engineering will be a highly sought-after profession. Creating an effective prompt comes with many challenges.
At one point, we felt that PlayBuddy's answers were too boring. We wanted its answers to be more fun - after all, we are an entertainment platform! So, our prompt specialist adjusted the instructions, telling PB to be “a bit more hilarious” in its responses. As a result, PB started hallucinating and ignoring the provided dataset almost entirely. If you asked PlayBuddy what the most-played game on the Play123 network was, it would confidently pick a random game from the web and give you completely incorrect play statistics!
“We designed a conversational system that is efficient and empathetic. The experience is completely different from standard website support bots. PlayBuddy’s focus is on delivering personalised answers in real time, based on the player’s preferences and past gaming history”
Tine Koloini
Here’s another common challenge: if you are too restrictive with how the LLM interacts and uses your internal data, your AI assistant can become dumb. On the other hand, if you use too much open data, your assistant might come up with irrelevant responses that have nothing to do with your business. There is a fine line between the two, and getting the balance between personality and accuracy just right requires extensive (and documented) testing with small, incremental changes.
In our case, Retoba provided the final prompt after many days of testing and fine-tuning it on our data. Our feedback was essential to this process, as only we could confirm which answers were accurate. The prompt itself remains encrypted and hidden from us, which gives you an idea of its value.
Have you noticed any surprising or creative ways users are engaging with PlayBuddy?
The most fascinating part for us is seeing the creative ways users interact with PB. We had our own ideas about how people would use it and prepared our data accordingly. However, after a few weeks of real-world testing, it was clear that our users had different ideas. They weren’t interested in the data we thought they’d want, so we changed the dataset to better reflect how people were actually using it.
Some users don’t want to spend time clicking around to find the game title they already know. Instead, they just type “take me to Drift Boss”, and PB will give them a direct link.
“Surprise me” is a request that is almost as common as “recommend a classic.”
Other users request things PB cannot handle yet, such as “Turn on the dark website theme.”
Finally, PlayBuddy is also trained to handle playful interactions like “Will you marry me?”
In general, what we see shows us that some of our users use PB to relax after a long day and just want to chat (“What should I do? I am bored.”). Whether this is a good use of the electricity required for each interaction is another matter entirely!
Having now run PlayBuddy for several months across Play123, what insights have you gained about user behaviour and preferences that you couldn’t have discovered through traditional analytics?
Traditional analytics don’t reveal users’ intimate thoughts and preferences, but AI interactions do. We anonymously analysed all user input (with AI assistance, of course!) and we now have a general picture of where we should head next.
What content do users want that we currently lack? Which games are frustrating them? What issues are they experiencing on our platform that they haven’t told us about? These are all questions that AI interactions can answer.
Where do you think AI assistants like PlayBuddy are heading? Do you see them becoming a standard part of online games platforms?
The traditional model of users searching, browsing, and clicking is gradually giving way to real-time conversations, and this transition is now evident in gaming as well. PlayBuddy’s unique blend of advanced AI with a human touch is an evolution that shows game developers, publishers, and industry entrepreneurs a glimpse into the future of player engagement in casual gaming.
“Users turn to PlayBuddy when they're trying to find a game but can't remember the name. With PB, they can quickly describe the game they’re looking for, and it will figure out which game they mean”
Tine Koloini
We believe that integrating AI into websites is more than just a new feature. It’s a new way of thinking about how platforms interact with their user base. Gaming has always been about creating memorable player experiences, and combining technology with empathy delivers that. Beyond that, smart gaming platforms will incorporate AI assistants into their websites because of the incredibly valuable first-party insights that they won’t get from anywhere else.
If you had to make one bold prediction about the future of AI and games, what would it be?
We are moving toward a future of hyper-personalised gaming experiences.
Traditionally, casual games feature a fixed set of levels that are essentially identical for all players. Now, with AI, each level can be adapted and personalised in real-time, based on the individual player’s gaming patterns. This way, users will feel as though each game was tailor-made for them.
I also believe that we could see a shift from “gaming on demand” to “gaming by wish,” allowing and fulfilling personalised requests such as: “I'd like to play Candy Crush, but with flowers instead of candies.”
Further down the rabbit hole
Some useful news, views and links to keep you going until next time:
Join AI Gamechangers as we host talks and panels live on stage. On Wednesday, 8th October, in Helsinki, Finland, the first AI Gamechangers Summit will take place as part of the PG Connects Helsinki, Big Screen Gaming Summit, and Finnish Games Week line-up! There’ll be a party, a careers zone and more. Tickets are still available, and we’re open to speaker submissions.
Homa co-founder and former chief revenue officer Olivier Le Bas has launched a new “AI-first entertainment company” called Everyday, focusing on building mobile games.
Dr Tommy Thompson and team are planning the second AI and Games conference for 3rd-4th November at Goldsmiths, University of London.
Speaking of Dr Tommy Thompson: Disney and Universal Studios are suing AI image generator Midjourney, and Tommy does a great round-up on the AI and Games blog about what it all means, including for the games industry.
It’s Google DeepMind’s turn to experiment with how its AI model plays Pokémon games. A report on Gemini 2.5 Pro says it panics when its Pokémon are close to death.
Charismatic.ai has unveiled a platform that transforms written prompts into animated "microdramas" in just 15 minutes, targeting agencies and creators with serialised short-form content generation. The platform, developed with backing from Channel 4 and Aardman, uses advanced video models like Veo3 to create episodic animated content. We interviewed Guy Gadney, the CEO of Charismatic.ai and Charisma.ai, last year.
Back in February we chatted with Kelly Vero, and in this video you can see her dive into more depth about using AI in NAK3D, her company that uses AI to rapidly generate 3D assets.